Microduck Training Studio R1.5.20 verification — 2026-10-07

Scope: source implementation and portable tests only.
No robot commands, firmware writes, live training, or GPU policy inference were performed.

Portable Python coverage: 138 distinct tests passed.
- Existing baseline regression: 111 tests.
- Recorded-start profile scenarios: 6 additional tests (with multiple invalid-frame subcases).
- Recorded-start request/API/isolation/log checks: 8 additional tests.
- Recorded-start math/reset/readback/runtime lifecycle: 13 additional tests.
The baseline+profile+first7 API cases ran as one 124-test suite; the final8 API,
5 existing target-limit API and13 new runtime cases passed in a 26-test follow-up
following the final log/readback changes.

JavaScript syntax check passed. Both target-limit UI and recovery UI Node
DOM/HTTP fixtures passed, covering standard/recorded launch, missing-pose error,
source-frame and numerical readback output, stopping and downloading.
This is not a real browser, 4K, CSS, Viser, or WebGL rendering test.

Existing real upstream recovery tests require the pinned simulator dependencies;
5 such tests were skipped because those dependencies are unavailable here.
Torch/MuJoCo/mjlab/ONNX runtime and the user's ONNX weights were not installed
or downloaded just for this verification. Actual roll/stand outcomes remain
unknown until the user runs the new mode on the existing training machine.

The supplied real recording was parsed successfully: the first StandUp enable
uses the immediately preceding near-static sample; all14 named policy joints
were extracted without mouth, and the original model hash was retained.
Pure quaternion projection of its gravity input agreed within 9e-9 L2.
This is not evidence of a successful physics rollout.

Changes preserve original target mapping and startup fairness checks. Recorded
pose application keeps sampled backlash/bias, reuses the existing geometric
plane placement, and performs no hidden physics steps. A first-action mismatch
in encoder pose, body gravity, qpos-derived gravity consistency, or zero velocity
blocks the run. Nonfinite numbers are reported as invalid, never tolerated.

The request mode cannot be sent into training, queues, export, PT play, ordinary
ONNX, or cached probe operations. Original model, recipe, parameters and user
recordings remain separate from this Public source package.

Public package validation also covers ZIP CRC, the SHA256 manifest for every
shipped file, executable start-training.command mode, and clean extracted imports.
